18:12
Safety for Conversational AI
-
Video
-
Audio
-
Table of contents
-
Video
-
Audio
- Description
- Transcript
- Discussion
About the talk
There is a current divide between methods applied in commercial and research-based Conversational AI systems: While the former rely on hand-written rules and templates, the latter apply neural end-to-end methods for response generation. In this talk, I argue that neural response generation needs to be safe and controllable in order to be suitable for end-user applications.
To support these claims, I will draw upon empirical results from two recent large scale evaluations: The Amazon Alexa Challenge for open domain conversation, where my team reached 3rd place in 2017 and 2019; And the E2E NLG challenge for generating responses in task-based dialogue systems, which my team organised.
Finally, I will provide examples on how my team tackles these challenges to develop safer Conversational AI.
00:45 About Neural Generation
01:20 The role of templates
01:24 Safety of GPT-3
01:54 About language models
05:00 Two case studies
05:25 Task-based systems
07:20 Open-domain systems
09:10 Microsoft Tay Bot incident
09:54 Systems misbehaving
10:22 Reinforcing gender stereotypes
11:05 How to detect abuse
12:00 SOTA in abuse mitigation
13:36 One size does not fit all
14:16 ConvAI Workshop
15:32 ALANA AI
16:57 Announcements
About speaker
Verena Rieser is a professor at Heriot-Watt University where she leads research on Natural Language Generation and Spoken Dialogue Systems, and a co-founder of the AI company Alana. She was recently awarded a Senior Research Fellowship by the Royal Society and is PI of several funded research projects and industry awards.
View the profileDialogue is a two-way communication at the moment. I'm only speaking to myself. So I'm a multimodal Tyler person and which makes this even more yet, and what I interested in my research building researching, whether we can build the system using machine learning. And to what extent is machine learning methods actually suitable for Building Systems, which a customer and more. Recently. I'm looking into end-to-end neural degeneration and in particular, the safety of these techniques, and then let's talk, I will
discuss the opportunities and limitations of his neck. The new generation is actually really nice dream these days. I think probably all of you heard about the amazing things to PT 3 can do. In fact, it's so mainstream that even people outside of Academia, talk about it. Even people outside of our field years that we from MC Hammer praising, the exciting things, but at least we can do So, now the question is, why do most commercial system actually still use templates? And one of the answers to this question, is that systems, like to 53 orange days. So,
shortly after its release, it got bad headlines because it was sexist and racist, and it was all over. So, readily giving advice on very sensitive topics. Such as suicide, which is obviously I'm not right to give you a little bit more background. Check 53 is a large language model. Some of you might also have heard of birth. And these large, some of these large language models can be used to generate language. So that is cheap in the TV series. That's about 35.
Get full access to Summit's talks
Interested in topic “Artificial Intelligence and Machine Learning”?
You might be interested in videos from this event
Similar talks
Conference Cast
With ConferenceCast.tv, you get access to our library of the world's best conference talks.
